Readback — voice AI that gets the number right

Call centres, logistics dispatchers, insurance agents and hospital staff read reference numbers over the phone dozens of times a day. In noise, 5 sounds like 9, M like N, S like F. One wrong character means a container ships to the wrong port, a payment gets reversed, or medication reaches the wrong patient. People cope by asking "Bravo for B or Delta for D?" for every character — losing 15 to 20 seconds per number and 30 percent of operator throughput. Readback listens in the background and knows what shape a reference number must take. Container numbers satisfy the ISO 6346 check digit. IBANs give mod-97. Cards give Luhn. This constraint does not improve the microphone — it multiplies the error budget the system can tolerate by 15×, measured across 12 million simulated calls in eight accents. Two thirds of the time Readback silently corrects a misheard character and shows the difference on screen. When it cannot, it asks about one character — once — and goes quiet again. The transcript is not stored. What comes back is the confirmed number and a count of how many times the conversation was interrupted. Built solo in 24 days: FastAPI + PostgreSQL backend, React + TypeScript frontend with a live AudioWorklet-to-PCM16-to-WebSocket audio path, AssemblyAI Universal-3.5 Pro Streaming with keyterm biasing and LLM Gateway. Deployed with a full admin panel, immutable audit log, tri-lingual interface (English, Uzbek, Russian), and 180 automated tests. Live: https://readback-eight.vercel.app · Try the pipeline on shipped fixtures: /demo